Overview
Mohamedabul/Qwen2.5-3B-CyberSecurity-Instruct is a specialized 3.1 billion parameter Qwen2.5-based model, fine-tuned using 4-bit QLoRA with Unsloth for enhanced performance. It functions as an expert cybersecurity analyst and reverse engineer, designed to understand and process complex security information efficiently. The model was trained on an extensive, unfiltered corpus including the NVD CVE Database (2020-2025), over 45,000 real-world exploits from Exploit-DB, and MITRE CWE classifications, totaling ~187,700 structured instruction samples.
Key Capabilities
- Vulnerability Triage: Automatically generates structured reports for CVEs, including severity, attack vectors, and mitigation strategies.
- Exploit Reverse-Engineering: Analyzes raw exploit code (C, Python, Bash) to provide technical breakdowns of how exploits function and their targeted vulnerabilities.
- Attack Chain Reasoning: Combines CVE information with exploit code to produce step-by-step kill-chain analyses, detailing the progression from initial access to system compromise.
Performance & Evaluation
Evaluated against an unseen hold-out dataset, the model achieved a Perplexity of 7.61, indicating high confidence in security concepts. Its METEOR score of 0.4084 demonstrates effective semantic understanding and appropriate use of security synonyms, while ROUGE-1 (0.3496) and ROUGE-L (0.2044) scores confirm strong structural and sentence-level alignment with security researcher standards.
Good For
- Automating initial vulnerability assessments and reporting.
- Understanding the technical mechanics of exploits without manual reverse engineering.
- Developing comprehensive attack chain analyses for incident response or threat intelligence.